The potential crosstalk between tumor and plasma cells and its association with clinical outcome and immunotherapy response in bladder cancer

The processed log₂(count + 1) bulk RNA-seq data (n = 430), MuTect2-processed somatic mutation data and corresponding clinical information of bladder urothelial carcinoma (BLCA) patients from The Cancer Genome Atlas (TCGA) were obtained using the UCSC Xena browser (https://​xenabrowser.​net/​datapages/​). RNA-seq data of the anti-PD-L1 therapy cohort (IMvigor210, n = 298) and sample annotation information, including binary best overall response (BOR), tumor mutation burden (TMB), tumor neoantigen burden (TNB), and survival information, were obtained using the IMvigor210CoreBiologies R package [22]. IMvigor210 was a single arm phase II study in patients with metastatic urothelial carcinoma (NCT02108652, NCT02951767) treated with atezolizumab (1200 mg 3 weekly). The primary endpoint was a partial response rate (PR) of more than 10% (RECIST v1.1). Raw scRNA-seq FASTQ data from two low-grade and six high-grade BC samples provided by the Union Hospital of Tongji Medical College, Wuhan, China (PRJNA662018) [23], were obtained from the Sequence Read Archive (SRA). The FASTQ data were filtered and read aligned using CellRanger (v.3.0.1, 10 × Genomics) [24], and feature barcode unique molecular identifier (UMI) matrices were generated based on the human reference genome GRCh38. Visium spatial transcriptomics data (GSE171351) [11] including h5ad files and raw were images downloaded from the Gene Expression Omnibus (GEO) and processed and analyzed using the Python package Scanpy (v.1.9.1) [25]. Briefly, spatial coordinates with total counts greater than 20,000, number of expressed genes greater than 6,000, and mitochondrial genes accounting for greater than 10% were filtered. The "normalize_total" function was used to normalize the counts and spatial information data, and the "highly_variable_genes" function was used to extract the top 2000 highly variable genes, followed by dimension reduction and clustering through principal component analysis (PCA), uniform manifold approximation and projection (UMAP) and Leiden algorithm analysis.

To assess the infiltration of immune cells in patients with BC, cell infiltration analysis was performed on bulk RNA-seq data using the R package IOBR [26] with the xCell method [27]. xCell is a method for scoring the abundances of 64 cell types based on single-sample gene set enrichment analysis (ssGSEA). Adaptive and innate immune cells were selected for further analysis. The list of marker genes in cells used in the calculations was downloaded from the study of Aran et al. [27]. Next, univariate Cox regression analysis was performed based on cell scores and patient survival data via the R package survival to predict the effect of cell infiltration scores on overall survival (OS), and Cox P < 0.05 was considered significantly associated with patient survival.

Cells were first filtered using the R package Seurat [28] using the following criteria: the number of expressed genes was lower than 300 or larger than 6000, and cells with more than 10% of UMIs mapped to mitochondrial genes were excluded. Only genes that were expressed in at least five cells were retained. After that, the data were normalized, and the top 2000 highly variable genes were detected by the "FindVariableFeatures" function. Next, the dimension of the scRNA-seq data was reduced by PCA based on 2000 genes, and 40 principal components were selected for subsequent analysis. Moreover, to remove the batch effects between samples, soft k-means clustering was performed using the R package harmony [29]. Finally, the "FindClusters" function was adopted for cell clustering, and the resolution was set to 0.5. To determine the cell types, we annotated the cells based on the CellMarker database [30] and corresponding literature.

Based on the cell annotation, PCs were extracted by the "subset" function, and dimension reduction and clustering were performed according to the same procedure as above, with resolution set to 0.4. To further identify IgG1- and IgA1-dominant cell subclusters, we classified cells according to the median expression of IGHG1 and IGHA1. Trajectory analysis was performed using Monocle 2 [31] to understand the transformation relationships between different types of PCs. Finally, to reveal the effects of IgG1 and IgA1 PCs on patient survival, an algorithm based on cosine similarity was first adopted using the R package COSG [32] to obtain the top 100 marker genes (Additional file 2: Table S1) of IgG1 or IgA1 PCs. The TCGA-BLCA samples were then grouped by the BC subtypes reported by Kamoun et al. [33], and ssGSEA was performed to estimate the cell abundance score for each patient using the R package GSVA based on the top 10 marker genes from the top 100 genes. Samples were grouped using the best cutoff or median based on cell abundance score, and survival analysis was performed using the R package survminer.

To identify malignant cells with clonal large-scale chromosomal copy number variations (CNVs), the CNVs of each cell were inferred using the R package infercnv [34]. We calculated the CNV score for each cell using the method of Peng et al. [35]. To determine the cell state, a nonnegative matrix factorization (NMF) algorithm was first employed for dimension reduction of all tumor cells, and subsequently, cells were scored for state using the R package AUCell [36] based on the gene sets of 12 cell states reported by Barkley et al. [8]. Finally, the R package COSG was used to screen all marker genes of cell state clusters (top 100) (Additional file 3: Table S2), and Gene Ontology (GO) [37], Kyoto Encyclopedia of Genes and Genomes (KEGG) [38], WikiPathways (WP) [39] and Reactome (REAC) [40] enrichment analyses were performed using g:Profiler [41]. The P value was adjusted by the g:SCS method.

To explore the crosstalk pattern between cells, we employed the R package CellChat [42]. Briefly, we followed the official workflow and created the CellChat object based on a normalized count matrix. Subsequently, the "identifyOverExpressedGenes" and "identifyOverExpressedInteractions" functions were used for data processing with a standard parameter set. The potential ligand/receptor (L/R) interactions among all cells, especially the interactions between PCs and tumor cells, were calculated and analyzed based on the functions "computeCommunProb", "computeCommunProbPathway", and "aggregateNet" using standard parameters.

To quantify the impact of communication between PCs and tumor cells on patient survival, we performed stepwise Cox regression of all L/R pairs (Additional file 4: Table S3) between PCs and tumor cells identified by CellChat using the R package My.stepwise, and the best multivariate Cox regression model based on TCGA-BLCA data was selected based on the Akaike information criterion. Next, we extracted the model coefficient (coef) of each L/R pair and calculated the risk score using the following formula:

The median risk score or optimal cutoff values were used to divide the patients into low- and high-risk groups. OS, disease-specific survival (DSS), disease-free survival (DFS), and progression-free survival (PFS) were analyzed using the Kaplan‒Meier method. The R package timeROC was used to assess the predictive power of the LRscore and clinical indicators of patient survival over time. The R package pROC was used to perform ROC analysis for the LRscore and selected anti-PD-L1 treatment-related indicators to compare their ability to discriminate between responding (R) and nonresponding (NR) patients. Multi-indicator ROC assessment was performed using binary logistic regression with linear fitting of multiple indicators based on SPSS (v.25.0).

To assess the association between the LRscore and cancer progression, we first calculated the cancer hallmark score for each sample based on the hallmark gene sets from MSigDB (http://​www.​gsea-msigdb.​org/​) using gene set variation analysis (GSVA) and subsequently calculated the correlation coefficient between the LRscore and cancer hallmark score using Spearman correlation analysis. The immune cell infiltration, immune, stromal, and ESTIMATE scores [43], tumor immune dysfunction and exclusion (TIDE) score [44], immunophenoscore (IPS) [45] and associated indicators were calculated using the R package IOBR [26], and their correlations with the LRscore were also evaluated by Spearman correlation analysis.

The classic subtype of BC reported by Robertson et al. [46] and the newest subtype reported by Kamoun et al. [33] were compared with our LRscore risk groups using the R package ggalluvial. The mutation profiles of the two risk group samples were analyzed using the R package maftools [47], and the significant differences between the two groups were evaluated using the chi-square test. To assess the response to targeted drugs in different risk groups, model training was performed using the R package oncoPredict [48] based on targeted drug-treated cell expression data from the GDSC database [49] to infer IC50 values for patients in the low- and high-risk groups. A low IC50 value generally indicates that a patient is more sensitive to a drug.

Statistical analysis and visualization were performed using R (v.4.1.1), Python (v.3.10), and Sangerbox [50]. The Wilcoxon test was used to compare two groups because the data were not normally distributed. The Kruskal–Wallis test was used to compare three or more groups. The P value was adjusted using the Benjamini and Hochberg method, if necessary. The detailed tools, methods, and thresholds for each analysis are described in the Materials and Methods section or the figure legends.

Inferential analysis of cellular infiltration (see Materials and Methods 2.2) based on TCGA-BLCA bulk RNA-seq data revealed that 11 immune cell types had a significant association with OS in the univariate Cox regression model (Additional file 1: Figure S1A), with high infiltration of CD8⁺ T_em cells, CD8⁺ T cells, NK cells, Th1 cells, and CD8⁺ T_cm cells associated with better OS (Cox P < 0.05), whereas high infiltration of four cell types (mast cells, neutrophils, M2 macrophages, and endothelial cells) was associated with worse OS (Cox P < 0.05). Among all B-cell types, only class-switched memory B cells (Cox P = 0.0059; log-rank P = 0.00054; HR = 0.77) and PCs (Cox P = 0.017; log-rank P = 0.019; HR = 0.82) were significantly associated with better OS (Fig. 2A and B). In the anti-PD-L1-treated IMvigor210 cohort, high infiltration of γδT (Tgd) cells, Th2 cells, and basophils was associated with better OS (Cox P < 0.05) (Additional file 1: Figure S1B). Among all B cells, only PCs were associated with better OS (Cox P = 0.0023; log-rank P = 0.0003; HR = 0.72) (Fig. 2C and D). Additionally, immunotherapy-responsive patients had a relatively higher PC abundance (Wilcoxon test, P < 0.05) (Fig. 2E). Overall, these bulk sequencing data-based cellular inference results suggest that high overall PC infiltration in BC patients predicts better OS and response to immunotherapy.

To further explore the heterogeneity of PCs at the single-cell level, we included an scRNA-seq dataset containing two low-grade and six high-grade BC samples. After quality control and cell filtering, 41,894 cells were obtained, including 30,180 epithelial cells (KRT18, EPCAM), 7,277 stromal cells (VWF, COL1A1), and 4,437 immune cells (PTPRC, CD19, MS4A2) (Fig. 3A and B). Among them, Cluster 11 was identified as PCs because of the high expression of CD79A, MZB1 and other PC markers (SDC1, CD79B, CD52, IGHG1/3/4, DERL3, FKBP11, JCHAIN, and CD38) (Fig. 3C and E). By comparing the proportion of cells between groups, we found a decrease in PCs in high-grade patients compared with low-grade patients (Fig. 3D, upper panel). Next, PCs were further subdivided into seven clusters (Fig. 3D, lower panel) without expression of naive B-cell markers, indicating that these cells were all mature PCs (Fig. 3E). Cluster 4 was identified as plasmablasts due to high expression of the proliferation-related gene MKI67, while the other cells mainly expressed IGHG1 and IGHA1 (Fig. 3F). Concurrently, other antibody genes (IGHG2, IGHA2, IGHM and IGHD) had low or almost no expression (Additional file 1: Figure S2A). After antibody gene-based dimension reduction through PCA, we observed that cells highly expressing IGHG1 and IGHA1 constituted almost all PCs, and there was overlap of cells with multiple IgG subtypes [51] (Additional file 1: Figure S2B). The scatter plot also showed a significant distinction between the two cell types (Fig. 3G), suggesting that IgG1 and IgA1 PCs were the dominant PC types in BC.

To further explore their functional characteristics, cells with high IGHG1 and low IGHA1 expression were defined as IgG1 PCs, and those with low IGHG1 and high IGHA1 expression were defined as IgA1 PCs (Additional file 1: Figure S3). Among them, IgA1 PCs were mainly involved in cell cycle-related pathway regulation (e.g., cell cycle and G1/S transition), while IgG1 PCs were significantly associated with immune responses (e.g., immune system and IL-17 signaling pathway) (Fig. 3G), suggesting that IgG1 PCs may be more likely to be involved in antitumor immune regulation. We further determined the order of differentiation of PCs by pseudotime trajectory analysis; IgA1 PCs were located at the beginning of the trajectory, and IgG1 PCs were located mainly at the end (Fig. 3H). Concurrently, we also observed that IGHA1 expression decreased at the end of the trajectory, while IGHG1 increased (Fig. 3I), suggesting a possible class-switched relationship between IgA1 and IgG1 PCs. Interestingly, the differentiation trajectory showed a clear branch at root 3 (Additional file 1: Figure S4), with cells of cell fate 2 presenting high expression of HSP family members and MHC-II-like molecules, which were absent in cell fate 1, implying a difference in the function of PCs with the two different fates.

To explore the effect of IgA1 and IgG1 PCs on cancer progression and patient survival, cell ratio analysis was performed; the results showed an increase in IgA1 PCs and a decrease in IgG1 PCs in patients with high-grade disease (Fig. 3J). However, heterogeneity was observed in the impact of IgA1 and IgG1 PCs on patient survival in different BC subtypes (Additional file 1: Figure S5A). For example, high signature scores of IgG1 predicted better OS in the basal/squamous (Ba/Sq) subtype (log-rank P = 0.0011) and luminal unstable (LumU) subtype but worse OS in the luminal papillary (LumP) subtype (log-rank P = 0.0075). In contrast, high signature scores of IgA1 PCs predicted better OS in the Ba/Sq subtype (log-rank P = 0.006) but tended to predict worse OS in the LumP and LumU subtypes. Although different types of PCs may have different functional roles in different subtypes, IgG1 PCs were dominant in all subtypes (Additional file 1: Figure S5B).

Based on copy number inference analysis in all single cells, tumor cells came from epithelial cells with significant copy number variation (CNV) (Fig. 4A), and almost all had amplifications in chromosomes 1, 8, 12, 17 and 19 and deletions in chromosome 14. Consistent with previous reports [23], cells from high-grade patients had higher CNV scores (Fig. 4B). Six tumor cell states (cycle, basal, hypoxia, partial epithelial-mesenchymal transition [pEMT], interferon, and stress) in BC were identified and used to assess the intrinsic functional characteristics of tumor cells (Fig. 4C and D). The tumor cells in different states were found to be involved in different regulatory functions (Fig. 4D). For example, cycle-like tumor cells were mainly associated with cell cycle checkpoints, and they may be involved in the T-cell receptor signaling pathway and regulate the PD-L1/PD-1 checkpoint pathway in cancer. Stress-like tumor cells were found to regulate the response to stress and the p53/IL-17 signaling pathway Additionally, it is worth noting that interferon-like tumor cells were found to be involved in the regulation of key immune response processes, such as antigen binding, MHC class II protein complex binding, immunoglobulin receptor binding, and PD-1 signaling. Although interferon-like tumor cells had the lowest cell proportion, they had the highest CNV scores (Fig. 4E).

We then evaluated the impact of different tumor cell states on patient survival (Additional file 1: Figure S6). In the nonimmunotherapy TCGA-BLCA cohort, high basal (Cox P = 0.015; HR = 1.20, 95% CI 1.00–1.40) and pEMT scores (Cox P = 0.044; HR = 1.20, 95% CI 1.00–1.40) were associated with worse OS. In the anti-PD-L1-treated IMvigor210 cohort, high cycle (Cox P = 0.012; HR = 0.84, 95% CI 0.73–0.96) and stress scores (Cox P = 0.004; HR = 1.30, 95% CI 1.10–1.50) were correlated with better and worse OS, respectively.

We next investigated the communication between all the cells. Clusters 10, 12, and 17 were further defined as myeloid-derived cells (MDCs) (LYZ, C1QB), T cells (CD3D, CD3E), and mast cells (TPSAB1 and TPSB2) (Additional file 1: Figure S7A). Among cells, stromal cells and MDCs were generally the strongest signaling senders and receivers, respectively (Additional file 1: Figure S7B). PCs were more likely to receive signals from stromal cells and send signals to T cells and MDCs (Additional file 1: Figure S7C), suggesting that PCs may be important immune mediators regulating tumor microenvironment (TME) intercellular communication in BC.

In the analysis of pathways based on L/R pairs between tumor cells and PCs, midkine (MK) and macrophage migration inhibitory factor (MIF) signaling were the most significantly enriched pathways (Additional file 1: Figure S8A and B). We further found that MK signaling was mainly enriched in tumor cells but not in PCs. In the MIF signaling pathway, tumor cells were the main output cells, especially interferon-like tumor cells (which were also the main signaling cells in all pathways), whereas PCs were the main receiver cells (Additional file 1: Figure S8C–E). Furthermore, it is worth noting that PCs were more likely to receive signals from interferon-like tumor cells and send signals to cycle-like tumor cells in low-grade BC samples (Additional file 1: Figure S9), while this communication pattern was disrupted in high-grade samples, in which PCs received additional signals from pEMT- and hypoxia-like tumor cells and sent more signals to interferon-like tumor cells. Together, these data suggest that the communication pattern between tumor cells and PCs dynamically changes with BC progression.

In the analysis of L/R pairs, the L/R pairs of MIF/(CD74 + CD44 or + CXCR4) and APP/CD74 were the most prominent interactions involving signal transduction from tumor cells to PCs (Fig. 5A). The protumor effects of MIF and APP have been reported in several studies, and their potential association with PCs may explain the additional mechanisms of tumor progression [52‐54]. Moreover, in the analysis of the correlation between L/R pairs and patient OS based on the anti-PD-L1 treatment cohort, some L/R pairs mediated signal transduction from hypoxia-like and stress-like tumor cells to IgG1 PCs (such as ANGPTL4/SDC1 and LAMB3/CD44, respectively), demonstrating a significant association with worse patient OS (log-rank P < 0.05), suggesting that these L/R pairs may play specific roles in patients receiving ICB therapy (Additional file 1: Figure S10A and S10B). On the other hand, some chemokine signaling-related molecules (such as CXCL10 and CXCL11) were associated with better patient OS (Fig. 5A), and these chemokines secreted by pEMT-like tumor cells may promote the recruitment of IgG1 PCs via ACKR1 [55]. In addition, the expression levels of 58 L/R pairs involved in significant communication between tumor cells and PCs were further investigated. In the TCGA-BLCA cohort, 33 molecules were differentially expressed between tumor and normal samples (FDR < 0.05), 64% of which were upregulated in tumor samples (Fig. 5B and C). In the anti-PD-L1 treatment IMvigor210 cohort, 16 molecules were differentially expressed between treatment-responsive and nonresponsive patients, but only 40% were upregulated in responsive patients (Fig. 5B and C). Thus, we speculated that global crosstalk between tumor cells and PCs may promote cancer development and be related to a worse response to ICB therapies.

To further select reliable L/R pairs between tumor cells and PCs, we assessed their expression levels in tissue sections using spatial transcriptome analysis. Sample GSM5224027 was used for expression assessment. We observed that Cluster 6 was located in a tertiary lymphoid-like structure in BC tissue, and the PC markers IGHG1 and IGHA1 also tended to be enriched (Additional file 1: Figure S11). Concurrently, the epithelial cell marker KRT18 was also found to be significantly expressed, whereas the expression of EPCAM was slightly lower. Based on the expression and biological localization analyses of L/R pairs, we identified two L/R pairs (LAMB3/CD44 and ANGPTL4/SDC1) with high reliability between tumor cells and PCs (Fig. 5D and Additional file 1: Figure S11). However, we found low ANGPTL4 expression in the tissue, which was also validated in the additional sample GSM5224029 (Fig. 5D and Additional file 1: Figure S12), suggesting that ANGPTL4 may be secreted only by some specific tumor cells. We examined the expression of ANGPTL4 in different cell types and found that it was significantly expressed in epithelial (tumor) cells and to a lesser extent in some stromal cells (Fig. 5E). In the analysis of the sample sources, a significant sample bias of ANGPTL4 was observed in samples B5 and B7, which could be related to their significant hypoxic levels (Fig. 5F and G). Hypoxic conditions induce ANGPTL4 expression [56]. We speculate that hypoxic tumor cells may secrete ANGPTL4 to communicate with SDC1⁺ PCs, affecting their differentiation, survival, and/or antibody secretion.

We further quantified the crosstalk patterns between tumor cells and PCs and assessed their association with patient survival and immunotherapy. Based on the TCGA-BLCA cohort, we selected 13 molecules from 58 L/R pairs involved in significant communication between tumor cells and PCs to construct a risk model (Fig. 6A), and the risk score was calculated according to the following equation:

We found that patients in the high LR score group had significantly worse OS (log-rank P < 0.0001) in the univariate Cox analysis (Fig. 6B). Stage, T stage, and N stage were also significantly associated with OS in the univariate Cox analysis (Cox P < 0.05, HR > 2) (Fig. 6C), while the LRscore was the only independent predictor for patients with BC (Cox P = 7.30E-08, HR = 2.6) after removing the confounding factors in the multivariate Cox analysis (Fig. 6C). The area under the tdROC curve (AUC) for 1-, 3-, 5-, 7-, and 10-year survival showed that the LRscore was a more stable and effective indicator than stage, T stage, and N stage (Fig. 6D). Moreover, a high LRscore was also associated with worse DSS (log-rank P < 0.0001), DFS (log-rank P = 0.024), and PFS (log-rank P < 0.0001) in BC patients (Fig. 6E), suggesting that the LRscore also has the potential to predict patient relapse and progression.

We further evaluated the effect of the LRscore on patient survival in the anti-PD-L1 treatment cohort, in which a high LRscore also predicted worse OS (log-rank P < 0.0001) (Fig. 6F). Concurrently, fewer patients responded to treatment in the high-risk group (high-risk vs. low-risk = 13% vs. 32%) (Fig. 6G), suggesting that the LRscore may be associated with poorer patient response to ICB therapies. This finding was further validated by the significant negative correlations of the LRscore with TNB (P < 0.0001, Pearson’s r = −0.29) and TMB (P < 0.0001, Pearson’s r = −0.28) (Fig. 6H). Moreover, we found that the expression of the model molecules ITGA6, HSPG2, LAMB1, and SDC2 was lower in responding patients (Wilcoxon test, P < 0.05), and low expression of these molecules was associated with poorer OS (log-rank P < 0.05) (Fig. 6I and J). These model molecules may be the main contributors to the success of the model in predicting the immune response in the anti-PD-L1 therapy cohort. Finally, we evaluated the predictive effect of the LRscore versus other immunotherapy-related predictors (TNB, TMB, PD-L1, PD1, and PCs) (Fig. 6K). The combination of the LRscore with TNB and TMB obtained the highest AUC (0.78) with a diagnostic specificity of 85.7%, and thus could be used to avoid excessive immunotherapy in patients.

To better analyze the features of the LR risk model, we first evaluated the correlation between the LRscore and cancer hallmarks based on the TCGA-BLCA cohort (Fig. 7A). The LRscore was positively correlated with most cancer hallmarks (P < 0.05, Spearman's r > 0.1), such as epithelial-mesenchymal transition, angiogenesis, hedgehog signaling, and hypoxia. Immune cell infiltration analysis showed that the LRscore was negatively correlated (P < 0.05, Spearman's r < −0.1) with the majority of tumor- suppressive T-cell types (Fig. 7B) and the IPS (P = 7.2E-14, Spearman's r = −0.36) (Fig. 7E). The IPS is an immunogenicity evaluation index, and a higher IPS indicates better immunotherapy performance. However, the LRscore was positively correlated with endothelial cells and the stromal score but not with effector cells (ECs), immune checkpoints (CPs), or the global immune score (Fig. 7C and E). More importantly, a significant positive association between the LRscore and TIDE score (P = 9.4E-15, Spearman's r = 0.38) was also observed, and this association was mainly derived from exclusion (P = 6.8E-34, Spearman's r = 0.56) and CAFs (P = 2.6E-26, Spearman's r = 0.50) rather than dysfunction (P = 0.37, Spearman's r = −0.04) (Fig. 7D). These data suggest that an immunosuppressed state reflected by the LRscore may mainly result from resistance of the tumor stroma to immune infiltration.

In addition, patients in the low-risk group were mostly categorized into the classic Robertson-LumP and consensus Kamoun-LumP molecular subtypes of BC, whereas the high-risk group was associated with malignant subtypes such as the Ba/Sq, Robertson-Neuronal, and Kamoun-NE-like subtypes (Fig. 7F). Remarkably, the LumP subtype showed a high FGFR3 mutation frequency in the low-risk group (Chi-square test, P < 0.001) (Fig. 7G). Moreover, we performed a drug sensitivity analysis based on the GDSC drug database for different risk groups, and patients in the low-risk group showed higher sensitivity to FGFR3-targeted drugs (such as PD173074 and AZD4547) (Fig. 7H). FGFR3 could be a drug target for patients in the low-risk group. Overall, these data suggest that a high LRscore predicts high immune exclusion and low immunogenicity in patients, suggesting poorer immunotherapy response and OS, but patients with a low LRscore may have better immune and targeted therapy outcomes.